The problem is that when a vendor finances their customers, they can create the illusion of 'real' demand for their product, when most of the the end-users are only actually using something because it's cheap. When the vendor runs low on cash and starts requiring payment, the customer may not be able to afford it, taking both vendor and customer down, and leaving the end-users who have a real need, and were willing to pay sustainable prices without any options.
that makes sense to me in a conceptual sense
however inference is very profitable and plummeting in cost for a given point on the intelligence curve, and nvidia gpus can serve different models so they are protected post-buildout